The AI Learning Hub
Published 1 November 2026

The EU AI Act's Deadlines Moved. Here's What Didn't.

If you work anywhere near AI in Europe, you have probably heard some version of “the AI Act got delayed.” That is true, but it is only half the story. Some of the most talked-about obligations did move. Others have been in force for more than a year and did not move at all. If you only hear the first half, you can end up relaxed about the wrong things.

This post sets out what changed in 2026, what already applies, and what a sensible organisation does with the extra time. It is dated, as everything in this journal is: written on 1 November 2026, and worth re-checking against the official text before you rely on it.

What changed

The EU adopted a package of simplification measures, the “Digital Omnibus”, and its AI part became Regulation (EU) 2026/1744. It entered into force on 27 July 2026. The headline effect is on the high-risk rules, the most demanding part of the Act:

  • High-risk systems in sensitive uses (the Annex III list: hiring, education, credit scoring, access to essential services, law enforcement and similar) now face their main obligations from 2 December 2027.
  • AI built into products that are already regulated (medical devices, machinery, toys and other products covered by EU product-safety law) now face them from 2 August 2028.

Those obligations are the heavy ones: risk management, data governance, technical documentation, logging, human oversight, accuracy and robustness testing, and registration. The reason given for the delay was practical. The harmonised standards that tell companies how to meet these requirements were not ready in time, and regulators did not want to enforce rules nobody could yet prove they met.

What did not change

Two layers of the Act were already live before the omnibus, and they still are.

The prohibitions. Since February 2025, a short list of AI practices has simply been banned in the EU. Examples include social scoring by public authorities, AI that manipulates people in ways that cause significant harm, untargeted scraping of facial images to build recognition databases, and emotion recognition in workplaces and schools (with narrow exceptions). None of this was paused.

The general-purpose model rules. Since August 2025, providers of general-purpose AI models, the large models that power chat assistants and many products, have had transparency and documentation duties, with extra obligations for the most capable models. These were not paused either.

The penalties also did not change. The maximum for the most serious breaches is still up to €35 million or 7% of worldwide annual turnover, whichever is higher.

What the delay is not

A delay is not a repeal, and it is not a signal that the requirements will be watered down further. The obligations themselves are largely the same; only their start dates moved. Organisations that stop preparing now will meet the same rules in 2027 or 2028 with less time to absorb them.

There is also a quieter cost to waiting. The high-risk requirements are mostly good engineering and governance practice: know what your system is for, test it before you ship it, keep records, and make sure a human can step in. Teams that already work this way will find compliance mostly a documentation exercise. Teams that don’t will find it a rebuild.

Outside the EU: the same pattern

Europe is not the only place where timelines are shifting. In the United States, Colorado passed the first broad state AI law in 2024. In 2026 it replaced that law with a narrower, transparency-focused one that takes effect on 1 January 2027. The direction is similar: lawmakers are keeping the goal of accountable AI while adjusting how much they demand, and when.

The lesson for anyone outside the legal team is simple. Treat AI regulation as a moving schedule around a fairly stable set of expectations. The dates will keep changing. The expectations — transparency, testing, human oversight, records — have barely moved in years.

What to do with the extra time

If your organisation builds or uses AI in any of the high-risk areas, the extra 15 to 24 months are best spent like this:

  • Make an inventory. List every AI system you build or use, what it decides or influences, and who is affected. You cannot classify what you have not found.
  • Classify honestly. For each system, ask whether it falls under a prohibition, the high-risk list, the general-purpose model rules, or none of them. Most systems will be “none”, and that is useful to know too.
  • Check the prohibitions today. These already apply. If anything you run comes close to the banned list, it is a current problem, not a 2027 one.
  • Start the habits, not the paperwork. Testing before release, logging, and a named human who can override the system are worth having anyway, and they are the hardest things to add late.
  • Watch the standards. The harmonised standards will define what “good enough” looks like. Follow them as they are published, rather than guessing.

Where to go deeper

Our free AI Essentials course covers the AI Act and other regulation in plain language, alongside how the technology works. If you are responsible for checking AI systems inside an organisation, the audit course walks through inventories, risk classification and evidence. Legal professionals can start with AI for Legal. It is orientation, not legal advice, and so is this post.

Related free courses